Hojun-Cho/ww
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

cgen: store the full register into padded scratch for a 3/5/6/7-byte aggregate tail (#10)

Browse Source 
The in-cap aggregate-receive materialise emitted a single narrow tail MOV that fell to MOVB for a 3/5/6/7-byte sub-8 tail, storing one byte while the scratch->dest copy read the full tail from uninitialised scratch — silently dropping members at the C2c whole-element arm (arr[i]=mk()) and loud-stopping at the #11 field arm. The scratch slot is ceil-8 padded (local_alloc/localadd round to 8) and the copy reads only tsz bytes, so flipping the tail default MOVB->MOVQ stores the full register harmlessly into the slot's own pad (in-bounds for in-cap <=24B); 1/2/4-byte tails stay byte-identical. Both stages symmetric. Removes the now-redundant #11 sub-8-tail loud-stop (keeps the float #165 and over-cap #234 loud-stops). The same narrow-tail materialise recurs at 6 other cstage sites (task #14).

Retires the obsolete idx_dot_aggret_subtail_loud //ww:error fixture (both stages now compile the case) and converts it to a positive cstage run-test; the struct-field shape is byte-id-divergent only via the pre-existing #9 frame-size bug, so the value pin uses array-field shapes. Value-asserting, reddens under each stage's independent revert.
This commit is contained in:
Hojun-Cho
2026-06-27 23:44:41 +09:00
parent 5b33ea3f7d
commit b3f4990979
5 changed files with 145 additions and 71 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
cmd/w6c/cgen.c
View File

@@ -5866,29 +5866,16 @@ cgexpr(Cg *c, Node *n, Local *locals)
"arr[i].f=mk() unwired (SSE "
"return eightbyte; #171)");
if (fsz > 8) {
/* A 3/5/6/7-byte sub-8 tail cannot be
* materialised by the single narrow MOV
* below — it stores ONE byte while the
* copy reads the full tail, dropping the
* rest from uninitialised scratch (a
* silent both-stage drop; empirically a
* 14B 7xi16 field loses f/g). The whole-
* element C2c sibling (:6764) shares this
* single-tail gap; until a general
* register->scratch tail (shift cascade)
* lands across BOTH sites, LOUD-STOP
* rather than silently drop — rule 7, the
* sibling of the over-cap/float stops
* above. Tails 0/1/2/4 are exact and flow
* through. Mirrors wwstage cgenexpr.ww. */
int tl11 = fsz % 8;
if (tl11 == 3 || tl11 == 5
|| tl11 == 6 || tl11 == 7)
fatal("#11: aggregate field receive "
"arr[i].f=mk() with a 3/5/6/7-byte "
"sub-8 tail unwired (materialise "
"single-MOV under-stores; "
"C2c-shared)");
/* The sub-8 tail materialise stores the FULL
* 8-byte register (MOVQ) into a ceil-8-padded
* scratch (cg_tagscr_slot -> local_alloc rounds
* to 8): the over-stored high bytes land in the
* pad and the scratch->dest copy reads only fsz
* bytes, so every in-cap tail (incl. 3/5/6/7) is
* exact without an immediate-shift cascade (w6a
* has no SHRQ $imm). Shares the C2c whole-element
* materialise (:7008). #10; mirrors wwstage
* cgenexpr.ww. */
int scr11 = cg_tagscr_slot(c,
&locals, fsz);
cgexpr(c, n->rhs, locals);
@@ -5903,10 +5890,11 @@ cgexpr(Cg *c, Node *n, Local *locals)
amem(D_BP, scr11
+ i11 * 8));
if (tail11 > 0) {
int op11 = (tail11 == 4)
? A_MOVL : (tail11
int op11 = (tail11 == 1)
? A_MOVB : (tail11
== 2) ? A_MOVW
: A_MOVB;
: (tail11 == 4)
? A_MOVL : A_MOVQ;
ins2(c, op11,
areg(regs11[full11]),
amem(D_BP, scr11
@@ -7015,8 +7003,9 @@ cgexpr(Cg *c, Node *n, Local *locals)
ins2(c, A_MOVQ, areg(regs[i]),
amem(D_BP, scr + i * 8));
if (tail > 0) {
int op = (tail == 4) ? A_MOVL
: (tail == 2) ? A_MOVW : A_MOVB;
int op = (tail == 1) ? A_MOVB
: (tail == 2) ? A_MOVW
: (tail == 4) ? A_MOVL : A_MOVQ;
ins2(c, op, areg(regs[full]),
amem(D_BP, scr + full * 8));
}

44
selfhost/cmd/wcc/cgenexpr.ww
View File

@@ -9127,9 +9127,10 @@ fn cgassign(c: *cgen, n: *syntax.node) void = {
wi += 1;
};
if (tail > 0) {
let top: str = "MOVB";
if (tail == 4) { top = "MOVL"; }
else { if (tail == 2) { top = "MOVW"; }; };
let top: str = "MOVQ";
if (tail == 1) { top = "MOVB"; }
else { if (tail == 2) { top = "MOVW"; }
else { if (tail == 4) { top = "MOVL"; }; }; };
let treg: str = "AX";
if (full == 1) { treg = "DX"; }
else { if (full == 2) { treg = "CX"; }; };
@@ -10054,27 +10055,15 @@ fn cgassign(c: *cgen, n: *syntax.node) void = {
// would smash the next field (g at +12).
let tsz: i32 = fi.fsz;
if (tsz > 8) {
// A 3/5/6/7-byte sub-8 tail cannot
// be materialised by the single narrow
// MOV below — it stores ONE byte while
// the copy reads the full tail, dropping
// the rest from uninitialised scratch (a
// silent both-stage drop; empirically a
// 14B 7xi16 field loses f/g). The whole-
// element C2c sibling (:9100) shares this
// single-tail gap; until a general
// register->scratch tail (shift cascade)
// lands across BOTH sites, LOUD-STOP
// rather than silently drop — rule 7, the
// #11 sibling of the over-cap/float stops
// above. Tails 0/1/2/4 are exact and flow
// through. Mirrors cstage cgen.c.
let tl11: i32 = tsz % 8;
if (tl11 == 3 || tl11 == 5 || tl11 == 6 || tl11 == 7) {
let m11t: str = "#11: aggregate field receive arr[i].f=mk() with a 3/5/6/7-byte sub-8 tail unwired (materialise single-MOV under-stores; C2c-shared)\n";
os.write(2, m11t.ptr, m11t.len: u64);
os.exit(1);
};
// The sub-8 tail materialise stores the FULL
// 8-byte register (MOVQ) into a ceil-8-padded
// scratch (tagscradd -> localadd rounds to 8):
// the over-stored high bytes land in the pad and
// the scratch->dest copy reads only tsz bytes, so
// every in-cap tail (incl. 3/5/6/7) is exact
// without an immediate-shift cascade (w6a has no
// SHRQ $imm). Shares the C2c whole-element
// materialise (:9100). #10; mirrors cstage cgen.c.
let scr11: i32 = tagscradd(c, tsz);
cgexpr(c, n.rhs);
// AX/DX/CX -> scratch (C2c materialise)
@@ -10093,9 +10082,10 @@ fn cgassign(c: *cgen, n: *syntax.node) void = {
wi11 += 1;
};
if (tail11 > 0) {
let top11: str = "MOVB";
if (tail11 == 4) { top11 = "MOVL"; }
else { if (tail11 == 2) { top11 = "MOVW"; }; };
let top11: str = "MOVQ";
if (tail11 == 1) { top11 = "MOVB"; }
else { if (tail11 == 2) { top11 = "MOVW"; }
else { if (tail11 == 4) { top11 = "MOVL"; }; }; };
let treg11: str = "AX";
if (full11 == 1) { treg11 = "DX"; }
else { if (full11 == 2) { treg11 = "CX"; }; };

82
test/lang/aggregate_tail_test.ww Normal file
View File

@@ -0,0 +1,82 @@
// aggregate_tail_test — in-cap aggregate-receive sub-8 tail materialise (#10).
// An in-cap (<=24B) aggregate-returning CALL received into an INDEXED dest is a
// two-step: (1) materialise the AX/DX/CX return regs into a frame scratch, (2)
// word-copy scratch -> dest. Step 1's sub-8 TAIL stored a single narrow MOV
// picked by (tail==4)?MOVL:(tail==2)?MOVW:MOVB, so for tail in {3,5,6,7} it fell
// to MOVB: one byte written while the copy reads the FULL tail, so the high tail
// bytes stayed uninitialised => dropped members. Two arms shared the gap — the
// C2c whole-element arr[i]=mk() (SILENT both stages, byte-id-BLIND: both emitted
// IDENTICAL wrong asm) and the #11 field-of-indexed arr[i].f=mk() (LOUD-STOPped).
// The fix stores the FULL register (MOVQ) into the ceil-8-padded scratch; the
// over-stored bytes die in the pad and the copy reads only the real size.
// A repro element must be align-<=2 UNPADDED: [7]i16 = 14B keeps a real tail-6;
// an i64-bearing struct pads to 16B (tail 0) and never trips it. PRIMITIVE-only
// asserts (deref_callarg lineage): a dropped word fails. The C2c value assert is
// the SOLE tooth there (byte-id blind); the #11 case re-arms a compile-error
// loud-stop on revert. Distinct nonzero per-element values so a stale read (the
// fully-dropped members read uninitialised, 0 on a fresh frame) mismatches.
package aggregate_tail_test;
type s = struct { f: [7]i16, g: i16 }; // f at off 0 (14B tail-6), g at off 14
fn mk7() [7]i16 = { return [10i16, 20i16, 30i16, 40i16, 50i16, 60i16, 70i16]; };
fn mk11() [11]u8 = { return [1u8,2u8,3u8,4u8,5u8,6u8,7u8,8u8,9u8,10u8,11u8]; };
fn mk13() [13]u8 = { return [1u8,2u8,3u8,4u8,5u8,6u8,7u8,8u8,9u8,10u8,11u8,12u8,13u8]; };
fn mk15() [15]u8 = { return [1u8,2u8,3u8,4u8,5u8,6u8,7u8,8u8,9u8,10u8,11u8,12u8,13u8,14u8,15u8]; };
@test fn c2c_whole_element() void = {
let arr: [3][7]i16;
arr[1] = mk7();
assert(arr[1][0] == 10i16);
assert(arr[1][1] == 20i16);
assert(arr[1][2] == 30i16);
assert(arr[1][3] == 40i16);
assert(arr[1][4] == 50i16); // eb1 byte 0-1
assert(arr[1][5] == 60i16); // eb1 byte 2-3 — dropped by the buggy MOVB tail
assert(arr[1][6] == 70i16); // eb1 byte 4-5 — dropped (fully-uninit tooth)
};
@test fn field_of_indexed() void = {
let arr: [3]s;
arr[1].g = 999i16; // neighbour set first
arr[1].f = mk7();
assert(arr[1].f[0] == 10i16);
assert(arr[1].f[4] == 50i16);
assert(arr[1].f[5] == 60i16);
assert(arr[1].f[6] == 70i16);
assert(arr[1].g == 999i16); // tail MOVQ writes the padded SCRATCH, not g
};
// The STRUCT-field-of-indexed variant ({t14,pad} struct field) is NOT pinned here:
// #10's materialise is byte-id clean for it, but the [N]box local frame size
// diverges cs!=ww (task #9, slotsize-vs-natural) — benign (both stages correct)
// yet byte-id-RED, so it stays out of the T2 corpus. Positive cstage coverage of
// that shape lives in test/wcc/data/idx_dot_aggret_subtail_run (converted from the
// retired loud-stop tripwire).
@test fn tail3() void = {
let a: [2][11]u8;
a[1] = mk11();
assert(a[1][0] == 1u8);
assert(a[1][7] == 8u8);
assert(a[1][8] == 9u8); // tail byte 0 (MOVB wrote only this)
assert(a[1][9] == 10u8); // tail byte 1 — dropped
assert(a[1][10] == 11u8); // tail byte 2 — dropped
};
@test fn tail5() void = {
let a: [2][13]u8;
a[1] = mk13();
assert(a[1][0] == 1u8);
assert(a[1][8] == 9u8);
assert(a[1][12] == 13u8); // tail byte 4 — dropped
};
@test fn tail7() void = {
let a: [2][15]u8;
a[1] = mk15();
assert(a[1][0] == 1u8);
assert(a[1][8] == 9u8);
assert(a[1][14] == 15u8); // tail byte 6 — dropped
};

16
test/wcc/data/idx_dot_aggret_subtail_loud/case.ww
View File

@@ -1,16 +0,0 @@
//ww:error "3/5/6/7-byte sub-8 tail unwired"
// #11 tripwire: an in-cap aggregate-returning CALL into an aggregate field of
// an indexed element where the field has a 3/5/6/7-byte sub-8 tail (here 14B
// 7xi16, tail=6). The materialise's single narrow MOV stores only one tail byte
// while the copy reads the full tail -> a SILENT both-stage member drop. Both
// stages LOUD-STOP until a general register->scratch tail lands (rule 7,
// C2c-shared). SUCCESS = the loud-stop regressed to a silent miscompile.
package main;
type t14 = struct { a:i16,b:i16,c:i16,d:i16,e:i16,f:i16,g:i16 };
type s14 = struct { x: t14, pad: i16 };
fn mk() t14 = { return t14{a=1i16,b=2i16,c=3i16,d=4i16,e=5i16,f=6i16,g=7i16}; };
export fn main() i32 = {
let arr: [2]s14;
arr[1].x = mk();
return 0;
};

29
test/wcc/data/idx_dot_aggret_subtail_run/case.ww Normal file
View File

@@ -0,0 +1,29 @@
//ww:run
// #10 positive (converted from the retired idx_dot_aggret_subtail_loud tripwire):
// an in-cap aggregate-returning CALL into a 14B-tail-6 STRUCT field of an indexed
// element (arr[i].x = mk()). Pre-#10 both stages LOUD-STOPped this 3/5/6/7-tail
// materialise; #10 stores the FULL register into the ceil-8-padded scratch, so the
// over-stored bytes die in the pad and every member round-trips. main returns 0
// only if all members AND the neighbour pad survived (else a distinct nonzero).
// cstage-only RUN by harness design (T1): this shape's [N]s14 local frame size
// diverges cs!=ww (task #9, slotsize-vs-natural, benign — both stages compute
// correct values) so it cannot enter the T2 byte-id corpus. Reverting #10 re-arms
// the loud-stop -> this case fails to compile (rc!=0) -> reddens.
package main;
type t14 = struct { a:i16,b:i16,c:i16,d:i16,e:i16,f:i16,g:i16 };
type s14 = struct { x: t14, pad: i16 };
fn mk() t14 = { return t14{a=10i16,b=20i16,c=30i16,d=40i16,e=50i16,f=60i16,g=70i16}; };
export fn main() i32 = {
let arr: [2]s14;
arr[1].pad = 999i16;
arr[1].x = mk();
if (arr[1].x.a != 10i16) { return 1; };
if (arr[1].x.b != 20i16) { return 2; };
if (arr[1].x.c != 30i16) { return 3; };
if (arr[1].x.d != 40i16) { return 4; };
if (arr[1].x.e != 50i16) { return 5; };
if (arr[1].x.f != 60i16) { return 6; }; // tail byte 0-1
if (arr[1].x.g != 70i16) { return 7; }; // tail byte 2-3 (dropped pre-#10)
if (arr[1].pad != 999i16) { return 8; }; // tail MOVQ must not smash pad
return 0;
};